add files and evaluate

README.md

@@ -0,0 +1,43 @@
+---
+language:
+- de
+license: apache-2.0
+tags:
+- automatic-speech-recognition
+- mozilla-foundation/common_voice_8_0
+- de
+- robust-speech-event
+datasets:
+- mozilla-foundation/common_voice_8_0
+model-index:
+- name: XLS-R-1B - German
+  results:
+  - task: 
+      name: Automatic Speech Recognition 
+      type: automatic-speech-recognition
+    dataset:
+      name: Common Voice 8
+      type: mozilla-foundation/common_voice_8_0
+      args: de
+    metrics:
+       - name: Test WER
+         type: wer
+         value: 11.37
+       - name: Test CER
+         type: cer
+         value: 2.89
+  - task: 
+      name: Automatic Speech Recognition
+      type: automatic-speech-recognition
+    dataset:
+      name: Robust Speech Event - Dev Data
+      type: speech-recognition-community-v2/dev_data
+      args: de
+    metrics:
+       - name: Dev WER
+         type: wer
+         value: 31.16
+       - name: Dev CER
+         type: cer
+         value: 13.41
+---

added_tokens.json

@@ -0,0 +1 @@
+{"<s>": 191, "</s>": 192}

all_results.json

@@ -0,0 +1,14 @@
+{
+    "epoch": 2.0,
+    "eval_loss": 0.11697383970022202,
+    "eval_runtime": 1050.8284,
+    "eval_samples": 16007,
+    "eval_samples_per_second": 15.233,
+    "eval_steps_per_second": 1.904,
+    "eval_wer": 0.1117051786299574,
+    "train_loss": 0.27212924943281636,
+    "train_runtime": 112594.4054,
+    "train_samples": 436168,
+    "train_samples_per_second": 7.748,
+    "train_steps_per_second": 0.242
+}

eval.py

@@ -0,0 +1,125 @@
+#!/usr/bin/env python3
+from datasets import load_dataset, load_metric, Audio, Dataset
+from transformers import pipeline, AutoFeatureExtractor
+import re
+import argparse
+import unicodedata
+from typing import Dict
+
+
+def log_results(result: Dataset, args: Dict[str, str]):
+    """ DO NOT CHANGE. This function computes and logs the result metrics. """
+
+    log_outputs = args.log_outputs
+    dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
+
+    # load metric
+    wer = load_metric("wer")
+    cer = load_metric("cer")
+
+    # compute metrics
+    wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
+    cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
+
+    # print & log results
+    result_str = (
+        f"WER: {wer_result}\n"
+        f"CER: {cer_result}"
+    )
+    print(result_str)
+
+    with open(f"{dataset_id}_eval_results.txt", "w") as f:
+        f.write(result_str)
+
+    # log all results in text file. Possibly interesting for analysis
+    if log_outputs is not None:
+        pred_file = f"log_{dataset_id}_predictions.txt"
+        target_file = f"log_{dataset_id}_targets.txt"
+
+        with open(pred_file, "w") as p, open(target_file, "w") as t:
+
+            # mapping function to write output
+            def write_to_file(batch, i):
+                p.write(f"{i}" + "\n")
+                p.write(batch["prediction"] + "\n")
+                t.write(f"{i}" + "\n")
+                t.write(batch["target"] + "\n")
+
+            result.map(write_to_file, with_indices=True)
+            
+
+def normalize_text(text: str) -> str:
+    """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
+
+    # From https://huggingface.co/jonatasgrosman/wav2vec2-large-xlsr-53-german.
+    CHARS_TO_IGNORE = [",", "?", "¿", ".", "!", "¡", ";", "；", ":", '""', "%", '"', "�", "ʿ", "·", "჻", "~", "՞",
+                       "؟", "،", "।", "॥", "«", "»", "„", "“", "”", "「", "」", "‘", "’", "《", "》", "(", ")", "[", "]",
+                       "{", "}", "=", "`", "_", "+", "<", ">", "…", "–", "°", "´", "ʾ", "‹", "›", "©", "®", "—", "→", "。",
+                       "、", "﹂", "﹁", "‧", "～", "﹏", "，", "｛", "｝", "（", "）", "［", "］", "【", "】", "‥", "〽",
+                       "『", "』", "〝", "〟", "⟨", "⟩", "〜", "：", "！", "？", "♪", "؛", "/", "\\", "º", "−", "^", "ʻ", "ˆ"]
+    chars_to_ignore_regex = f"[{re.escape(''.join(CHARS_TO_IGNORE))}]"
+    text = re.sub(chars_to_ignore_regex, "", text.lower())
+
+    return " ".join(text.split())
+
+
+def main(args):
+    # load dataset
+    dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
+
+    # # for testing: only process the first two examples as a test
+    # dataset = dataset.select(range(10))
+
+    # load processor
+    feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
+    sampling_rate = feature_extractor.sampling_rate
+
+    # resample audio
+    dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
+
+    # load eval pipeline
+    asr = pipeline("automatic-speech-recognition", model=args.model_id, device=0)
+
+    # map function to decode audio
+    def map_to_pred(batch):
+        prediction = asr(batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s)
+
+        batch["prediction"] = prediction["text"]
+        batch["target"] = normalize_text(batch["sentence"])
+        return batch
+
+    # run inference on all examples
+    result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
+
+    # compute and log_results
+    # do not change function below
+    log_results(result, args)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
+    )
+    parser.add_argument(
+        "--dataset", type=str, required=True, help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets"
+    )
+    parser.add_argument(
+        "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'`  for Common Voice"
+    )
+    parser.add_argument(
+        "--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`"
+    )
+    parser.add_argument(
+        "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to None. For long audio files a good value would be 5.0 seconds."
+    )
+    parser.add_argument(
+        "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to None. For long audio files a good value would be 1.0 seconds."
+    )
+    parser.add_argument(
+        "--log_outputs", action='store_true', help="If defined, write outputs to log file for analysis."
+    )
+    args = parser.parse_args()
+
+    main(args)

eval_results.json

@@ -0,0 +1,9 @@
+{
+    "epoch": 2.0,
+    "eval_loss": 0.11697383970022202,
+    "eval_runtime": 1050.8284,
+    "eval_samples": 16007,
+    "eval_samples_per_second": 15.233,
+    "eval_steps_per_second": 1.904,
+    "eval_wer": 0.1117051786299574
+}

mozilla-foundation_common_voice_8_0_de_test_eval_results.txt

@@ -0,0 +1,2 @@
+WER: 0.11368385114373507
+CER: 0.028929306965087716

preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

run.sh

@@ -0,0 +1,33 @@
+python run_speech_recognition_ctc.py \
+	--dataset_name="mozilla-foundation/common_voice_8_0" \
+	--model_name_or_path="facebook/wav2vec2-xls-r-1b" \
+	--dataset_config_name="de" \
+	--output_dir="./" \
+	--overwrite_output_dir \
+	--num_train_epochs="2" \
+	--per_device_train_batch_size="8" \
+	--per_device_eval_batch_size="8" \
+	--gradient_accumulation_steps="4" \
+	--learning_rate="7e-5" \
+	--lr_scheduler_type="cosine" \
+	--warmup_steps="10" \
+	--length_column_name="input_length" \
+	--evaluation_strategy="steps" \
+	--text_column_name="sentence" \
+	--chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \
+	--save_steps="1500" \
+	--eval_steps="1500" \
+	--logging_steps="100" \
+	--save_total_limit="2" \
+	--freeze_feature_encoder \
+	--attention_dropout="0.15" \
+	--hidden_dropout="0.15" \
+	--mask_time_prob="0.15" \
+	--mask_time_length="10" \
+	--mask_feature_prob="0.35" \
+	--mask_feature_length="64" \
+	--gradient_checkpointing \
+	--use_auth_token \
+	--fp16 \
+	--group_by_length \
+	--do_train --do_eval

run_speech_recognition_ctc.py

@@ -0,0 +1,737 @@
+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+
+""" Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
+
+import functools
+import json
+import logging
+import os
+import re
+import sys
+import warnings
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+import datasets
+import numpy as np
+import torch
+from datasets import DatasetDict, load_dataset, load_metric
+
+import transformers
+from transformers import (
+    AutoConfig,
+    AutoFeatureExtractor,
+    AutoModelForCTC,
+    AutoProcessor,
+    AutoTokenizer,
+    HfArgumentParser,
+    Trainer,
+    TrainingArguments,
+    Wav2Vec2Processor,
+    set_seed,
+)
+from transformers.trainer_utils import get_last_checkpoint, is_main_process
+from transformers.utils import check_min_version
+from transformers.utils.versions import require_version
+
+
+# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
+check_min_version("4.17.0.dev0")
+
+require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
+
+
+logger = logging.getLogger(__name__)
+
+
+def list_field(default=None, metadata=None):
+    return field(default_factory=lambda: default, metadata=metadata)
+
+
+@dataclass
+class ModelArguments:
+    """
+    Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
+    """
+
+    model_name_or_path: str = field(
+        metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
+    )
+    tokenizer_name_or_path: Optional[str] = field(
+        default=None,
+        metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
+    )
+    cache_dir: Optional[str] = field(
+        default=None,
+        metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
+    )
+    freeze_feature_encoder: bool = field(
+        default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
+    )
+    attention_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
+    )
+    activation_dropout: float = field(
+        default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
+    )
+    feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
+    hidden_dropout: float = field(
+        default=0.0,
+        metadata={
+            "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
+        },
+    )
+    final_dropout: float = field(
+        default=0.0,
+        metadata={"help": "The dropout probability for the final projection layer."},
+    )
+    mask_time_prob: float = field(
+        default=0.05,
+        metadata={
+            "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
+            "vectors will be masked along the time axis."
+        },
+    )
+    mask_time_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the time axis."},
+    )
+    mask_feature_prob: float = field(
+        default=0.0,
+        metadata={
+            "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
+            "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
+        },
+    )
+    mask_feature_length: int = field(
+        default=10,
+        metadata={"help": "Length of vector span to mask along the feature axis."},
+    )
+    layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
+    ctc_loss_reduction: Optional[str] = field(
+        default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
+    )
+
+
+@dataclass
+class DataTrainingArguments:
+    """
+    Arguments pertaining to what data we are going to input our model for training and eval.
+
+    Using `HfArgumentParser` we can turn this class
+    into argparse arguments to be able to specify them on
+    the command line.
+    """
+
+    dataset_name: str = field(
+        metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    dataset_config_name: str = field(
+        default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
+    )
+    train_split_name: str = field(
+        default="train+validation",
+        metadata={
+            "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train+validation'"
+        },
+    )
+    eval_split_name: str = field(
+        default="test",
+        metadata={
+            "help": "The name of the evaluation data set split to use (via the datasets library). Defaults to 'test'"
+        },
+    )
+    audio_column_name: str = field(
+        default="audio",
+        metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
+    )
+    text_column_name: str = field(
+        default="text",
+        metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
+    )
+    overwrite_cache: bool = field(
+        default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
+    )
+    preprocessing_num_workers: Optional[int] = field(
+        default=None,
+        metadata={"help": "The number of processes to use for the preprocessing."},
+    )
+    max_train_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        },
+    )
+    max_eval_samples: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
+            "value if set."
+        },
+    )
+    chars_to_ignore: Optional[List[str]] = list_field(
+        default=None,
+        metadata={"help": "A list of characters to remove from the transcripts."},
+    )
+    eval_metrics: List[str] = list_field(
+        default=["wer"],
+        metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
+    )
+    max_duration_in_seconds: float = field(
+        default=20.0,
+        metadata={
+            "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
+        },
+    )
+    min_duration_in_seconds: float = field(
+        default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
+    )
+    preprocessing_only: bool = field(
+        default=False,
+        metadata={
+            "help": "Whether to only do data preprocessing and skip training. "
+            "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
+            "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
+            "so that the cached datasets can consequently be loaded in distributed training"
+        },
+    )
+    use_auth_token: bool = field(
+        default=False,
+        metadata={
+            "help": "If :obj:`True`, will use the token generated when running"
+            ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
+        },
+    )
+    unk_token: str = field(
+        default="[UNK]",
+        metadata={"help": "The unk token for the tokenizer"},
+    )
+    pad_token: str = field(
+        default="[PAD]",
+        metadata={"help": "The padding token for the tokenizer"},
+    )
+    word_delimiter_token: str = field(
+        default="|",
+        metadata={"help": "The word delimiter token for the tokenizer"},
+    )
+    phoneme_language: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "The target language that should be used be"
+            " passed to the tokenizer for tokenization. Note that"
+            " this is only relevant if the model classifies the"
+            " input audio to a sequence of phoneme sequences."
+        },
+    )
+
+
+@dataclass
+class DataCollatorCTCWithPadding:
+    """
+    Data collator that will dynamically pad the inputs received.
+    Args:
+        processor (:class:`~transformers.AutoProcessor`)
+            The processor used for proccessing the data.
+        padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
+            Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
+            among:
+            * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
+              sequence if provided).
+            * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
+              maximum acceptable input length for the model if that argument is not provided.
+            * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
+              different lengths).
+        max_length (:obj:`int`, `optional`):
+            Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
+        max_length_labels (:obj:`int`, `optional`):
+            Maximum length of the ``labels`` returned list and optionally padding length (see above).
+        pad_to_multiple_of (:obj:`int`, `optional`):
+            If set will pad the sequence to a multiple of the provided value.
+            This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
+            7.5 (Volta).
+    """
+
+    processor: AutoProcessor
+    padding: Union[bool, str] = "longest"
+    pad_to_multiple_of: Optional[int] = None
+    pad_to_multiple_of_labels: Optional[int] = None
+
+    def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
+        # split inputs and labels since they have to be of different lenghts and need
+        # different padding methods
+        input_features = [{"input_values": feature["input_values"]} for feature in features]
+        label_features = [{"input_ids": feature["labels"]} for feature in features]
+
+        batch = self.processor.pad(
+            input_features,
+            padding=self.padding,
+            pad_to_multiple_of=self.pad_to_multiple_of,
+            return_tensors="pt",
+        )
+
+        with self.processor.as_target_processor():
+            labels_batch = self.processor.pad(
+                label_features,
+                padding=self.padding,
+                pad_to_multiple_of=self.pad_to_multiple_of_labels,
+                return_tensors="pt",
+            )
+
+        # replace padding with -100 to ignore loss correctly
+        labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
+
+        batch["labels"] = labels
+
+        return batch
+
+
+def create_vocabulary_from_data(
+    datasets: DatasetDict,
+    word_delimiter_token: Optional[str] = None,
+    unk_token: Optional[str] = None,
+    pad_token: Optional[str] = None,
+):
+    # Given training and test labels create vocabulary
+    def extract_all_chars(batch):
+        all_text = " ".join(batch["target_text"])
+        vocab = list(set(all_text))
+        return {"vocab": [vocab], "all_text": [all_text]}
+
+    vocabs = datasets.map(
+        extract_all_chars,
+        batched=True,
+        batch_size=-1,
+        keep_in_memory=True,
+        remove_columns=datasets["train"].column_names,
+    )
+
+    # take union of all unique characters in each dataset
+    vocab_set = functools.reduce(
+        lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
+    )
+
+    vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
+
+    # replace white space with delimiter token
+    if word_delimiter_token is not None:
+        vocab_dict[word_delimiter_token] = vocab_dict[" "]
+        del vocab_dict[" "]
+
+    # add unk and pad token
+    if unk_token is not None:
+        vocab_dict[unk_token] = len(vocab_dict)
+
+    if pad_token is not None:
+        vocab_dict[pad_token] = len(vocab_dict)
+
+    return vocab_dict
+
+
+def main():
+    # See all possible arguments in src/transformers/training_args.py
+    # or by passing the --help flag to this script.
+    # We now keep distinct sets of args, for a cleaner separation of concerns.
+
+    parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
+    if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
+        # If we pass only one argument to the script and it's the path to a json file,
+        # let's parse it to get our arguments.
+        model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
+    else:
+        model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    # Detecting last checkpoint.
+    last_checkpoint = None
+    if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
+        last_checkpoint = get_last_checkpoint(training_args.output_dir)
+        if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
+            raise ValueError(
+                f"Output directory ({training_args.output_dir}) already exists and is not empty. "
+                "Use --overwrite_output_dir to overcome."
+            )
+        elif last_checkpoint is not None:
+            logger.info(
+                f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
+                "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
+            )
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        handlers=[logging.StreamHandler(sys.stdout)],
+    )
+    logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
+
+    # Log on each process the small summary:
+    logger.warning(
+        f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
+        f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
+    )
+    # Set the verbosity to info of the Transformers logger (on main process only):
+    if is_main_process(training_args.local_rank):
+        transformers.utils.logging.set_verbosity_info()
+    logger.info("Training/evaluation parameters %s", training_args)
+
+    # Set seed before initializing model.
+    set_seed(training_args.seed)
+
+    # 1. First, let's load the dataset
+    raw_datasets = DatasetDict()
+
+    if training_args.do_train:
+        raw_datasets["train"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.train_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.audio_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--audio_column_name` to the correct audio column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.text_column_name not in raw_datasets["train"].column_names:
+            raise ValueError(
+                f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
+                "Make sure to set `--text_column_name` to the correct text column - one of "
+                f"{', '.join(raw_datasets['train'].column_names)}."
+            )
+
+        if data_args.max_train_samples is not None:
+            raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
+
+    if training_args.do_eval:
+        raw_datasets["eval"] = load_dataset(
+            data_args.dataset_name,
+            data_args.dataset_config_name,
+            split=data_args.eval_split_name,
+            use_auth_token=data_args.use_auth_token,
+        )
+
+        if data_args.max_eval_samples is not None:
+            raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
+
+    # 2. We remove some special characters from the datasets
+    # that make training complicated and do not help in transcribing the speech
+    # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
+    # that could be easily picked up by the model
+    chars_to_ignore_regex = (
+        f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
+    )
+    text_column_name = data_args.text_column_name
+
+    def remove_special_characters(batch):
+        if chars_to_ignore_regex is not None:
+            batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
+        else:
+            batch["target_text"] = batch[text_column_name].lower() + " "
+        return batch
+
+    with training_args.main_process_first(desc="dataset map special characters removal"):
+        raw_datasets = raw_datasets.map(
+            remove_special_characters,
+            remove_columns=[text_column_name],
+            desc="remove special characters from datasets",
+        )
+
+    # save special tokens for tokenizer
+    word_delimiter_token = data_args.word_delimiter_token
+    unk_token = data_args.unk_token
+    pad_token = data_args.pad_token
+
+    # 3. Next, let's load the config as we might need it to create
+    # the tokenizer
+    # load config
+    config = AutoConfig.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # 4. Next, if no tokenizer file is defined,
+    # we create the vocabulary of the model by extracting all unique characters from
+    # the training and evaluation datasets
+    # We need to make sure that only first rank saves vocabulary
+    # make sure all processes wait until vocab is created
+    tokenizer_name_or_path = model_args.tokenizer_name_or_path
+    tokenizer_kwargs = {}
+    if tokenizer_name_or_path is None:
+        # save vocab in training output dir
+        tokenizer_name_or_path = training_args.output_dir
+
+        vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
+
+        with training_args.main_process_first():
+            if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
+                os.remove(vocab_file)
+
+        with training_args.main_process_first(desc="dataset map vocabulary creation"):
+            if not os.path.isfile(vocab_file):
+                os.makedirs(tokenizer_name_or_path, exist_ok=True)
+                vocab_dict = create_vocabulary_from_data(
+                    raw_datasets,
+                    word_delimiter_token=word_delimiter_token,
+                    unk_token=unk_token,
+                    pad_token=pad_token,
+                )
+
+                # save vocab dict to be loaded into tokenizer
+                with open(vocab_file, "w") as file:
+                    json.dump(vocab_dict, file)
+
+        # if tokenizer has just been created
+        # it is defined by `tokenizer_class` if present in config else by `model_type`
+        tokenizer_kwargs = {
+            "config": config if config.tokenizer_class is not None else None,
+            "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
+            "unk_token": unk_token,
+            "pad_token": pad_token,
+            "word_delimiter_token": word_delimiter_token,
+        }
+
+    # 5. Now we can instantiate the feature extractor, tokenizer and model
+    # Note for distributed training, the .from_pretrained methods guarantee that only
+    # one local process can concurrently download model & vocab.
+
+    # load feature_extractor and tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(
+        tokenizer_name_or_path,
+        use_auth_token=data_args.use_auth_token,
+        **tokenizer_kwargs,
+    )
+    feature_extractor = AutoFeatureExtractor.from_pretrained(
+        model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
+    )
+
+    # adapt config
+    config.update(
+        {
+            "feat_proj_dropout": model_args.feat_proj_dropout,
+            "attention_dropout": model_args.attention_dropout,
+            "hidden_dropout": model_args.hidden_dropout,
+            "final_dropout": model_args.final_dropout,
+            "mask_time_prob": model_args.mask_time_prob,
+            "mask_time_length": model_args.mask_time_length,
+            "mask_feature_prob": model_args.mask_feature_prob,
+            "mask_feature_length": model_args.mask_feature_length,
+            "gradient_checkpointing": training_args.gradient_checkpointing,
+            "layerdrop": model_args.layerdrop,
+            "ctc_loss_reduction": model_args.ctc_loss_reduction,
+            "pad_token_id": tokenizer.pad_token_id,
+            "vocab_size": len(tokenizer),
+            "activation_dropout": model_args.activation_dropout,
+        }
+    )
+
+    # create model
+    model = AutoModelForCTC.from_pretrained(
+        model_args.model_name_or_path,
+        cache_dir=model_args.cache_dir,
+        config=config,
+        use_auth_token=data_args.use_auth_token,
+    )
+
+    # freeze encoder
+    if model_args.freeze_feature_encoder:
+        model.freeze_feature_encoder()
+
+    # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
+    # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
+    # so that we just need to set the correct target sampling rate and normalize the input
+    # via the `feature_extractor`
+
+    # make sure that dataset decodes audio with correct sampling rate
+    dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
+    if dataset_sampling_rate != feature_extractor.sampling_rate:
+        raw_datasets = raw_datasets.cast_column(
+            data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
+        )
+
+    # derive max & min input length for sample rate & max duration
+    max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
+    min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
+    audio_column_name = data_args.audio_column_name
+    num_workers = data_args.preprocessing_num_workers
+
+    # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
+    phoneme_language = data_args.phoneme_language
+
+    # Preprocessing the datasets.
+    # We need to read the audio files as arrays and tokenize the targets.
+    def prepare_dataset(batch):
+        # load audio
+        sample = batch[audio_column_name]
+
+        inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
+        batch["input_values"] = inputs.input_values[0]
+        batch["input_length"] = len(batch["input_values"])
+
+        # encode targets
+        additional_kwargs = {}
+        if phoneme_language is not None:
+            additional_kwargs["phonemizer_lang"] = phoneme_language
+
+        batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
+        return batch
+
+    with training_args.main_process_first(desc="dataset map preprocessing"):
+        vectorized_datasets = raw_datasets.map(
+            prepare_dataset,
+            remove_columns=next(iter(raw_datasets.values())).column_names,
+            num_proc=num_workers,
+            desc="preprocess datasets",
+        )
+
+        def is_audio_in_length_range(length):
+            return length > min_input_length and length < max_input_length
+
+        # filter data that is shorter than min_input_length
+        vectorized_datasets = vectorized_datasets.filter(
+            is_audio_in_length_range,
+            num_proc=num_workers,
+            input_columns=["input_length"],
+        )
+
+    # 7. Next, we can prepare the training.
+    # Let's use word error rate (WER) as our evaluation metric,
+    # instantiate a data collator and the trainer
+
+    # Define evaluation metrics during training, *i.e.* word error rate, character error rate
+    eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
+
+    # for large datasets it is advised to run the preprocessing on a
+    # single machine first with ``args.preprocessing_only`` since there will mostly likely
+    # be a timeout when running the script in distributed mode.
+    # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
+    # cached dataset
+    if data_args.preprocessing_only:
+        logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
+        return
+
+    def compute_metrics(pred):
+        pred_logits = pred.predictions
+        pred_ids = np.argmax(pred_logits, axis=-1)
+
+        pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
+
+        pred_str = tokenizer.batch_decode(pred_ids)
+        # we do not want to group tokens when computing the metrics
+        label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
+
+        metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
+
+        return metrics
+
+    # Now save everything to be able to create a single processor later
+    if is_main_process(training_args.local_rank):
+        # save feature extractor, tokenizer and config
+        feature_extractor.save_pretrained(training_args.output_dir)
+        tokenizer.save_pretrained(training_args.output_dir)
+        config.save_pretrained(training_args.output_dir)
+
+    try:
+        processor = AutoProcessor.from_pretrained(training_args.output_dir)
+    except (OSError, KeyError):
+        warnings.warn(
+            "Loading a processor from a feature extractor config that does not"
+            " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
+            " attribute to your `preprocessor_config.json` file to suppress this warning: "
+            " `'processor_class': 'Wav2Vec2Processor'`",
+            FutureWarning,
+        )
+        processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
+
+    # Instantiate custom data collator
+    data_collator = DataCollatorCTCWithPadding(processor=processor)
+
+    # Initialize Trainer
+    trainer = Trainer(
+        model=model,
+        data_collator=data_collator,
+        args=training_args,
+        compute_metrics=compute_metrics,
+        train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
+        eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
+        tokenizer=feature_extractor,
+    )
+
+    # 8. Finally, we can start training
+
+    # Training
+    if training_args.do_train:
+
+        # use last checkpoint if exist
+        if last_checkpoint is not None:
+            checkpoint = last_checkpoint
+        elif os.path.isdir(model_args.model_name_or_path):
+            checkpoint = model_args.model_name_or_path
+        else:
+            checkpoint = None
+
+        train_result = trainer.train(resume_from_checkpoint=checkpoint)
+        trainer.save_model()
+
+        metrics = train_result.metrics
+        max_train_samples = (
+            data_args.max_train_samples
+            if data_args.max_train_samples is not None
+            else len(vectorized_datasets["train"])
+        )
+        metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
+
+        trainer.log_metrics("train", metrics)
+        trainer.save_metrics("train", metrics)
+        trainer.save_state()
+
+    # Evaluation
+    results = {}
+    if training_args.do_eval:
+        logger.info("*** Evaluate ***")
+        metrics = trainer.evaluate()
+        max_eval_samples = (
+            data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
+        )
+        metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
+
+        trainer.log_metrics("eval", metrics)
+        trainer.save_metrics("eval", metrics)
+
+    # Write model card and (optionally) push to hub
+    config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
+    kwargs = {
+        "finetuned_from": model_args.model_name_or_path,
+        "tasks": "speech-recognition",
+        "tags": ["automatic-speech-recognition", data_args.dataset_name],
+        "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
+        "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
+    }
+    if "common_voice" in data_args.dataset_name:
+        kwargs["language"] = config_name
+
+    if training_args.push_to_hub:
+        trainer.push_to_hub(**kwargs)
+    else:
+        trainer.create_model_card(**kwargs)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}

speech-recognition-community-v2_dev_data_de_validation_eval_results.txt

@@ -0,0 +1,2 @@
+WER: 0.31160744887467157
+CER: 0.1341142698398

tokenizer_config.json

@@ -0,0 +1 @@
+{"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}

train_results.json

@@ -0,0 +1,8 @@
+{
+    "epoch": 2.0,
+    "train_loss": 0.27212924943281636,
+    "train_runtime": 112594.4054,
+    "train_samples": 436168,
+    "train_samples_per_second": 7.748,
+    "train_steps_per_second": 0.242
+}

trainer_state.json

@@ -0,0 +1,1798 @@
+{
+  "best_metric": null,
+  "best_model_checkpoint": null,
+  "epoch": 1.9809064397204748,
+  "global_step": 27000,
+  "is_hyper_param_search": false,
+  "is_local_process_zero": true,
+  "is_world_process_zero": true,
+  "log_history": [
+    {
+      "epoch": 0.01,
+      "learning_rate": 6.999815761247666e-05,
+      "loss": 4.5119,
+      "step": 100
+    },
+    {
+      "epoch": 0.01,
+      "learning_rate": 6.999169171843241e-05,
+      "loss": 2.6638,
+      "step": 200
+    },
+    {
+      "epoch": 0.02,
+      "learning_rate": 6.998057503298084e-05,
+      "loss": 1.183,
+      "step": 300
+    },
+    {
+      "epoch": 0.03,
+      "learning_rate": 6.996480903365516e-05,
+      "loss": 0.7325,
+      "step": 400
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 6.99443958159349e-05,
+      "loss": 0.5966,
+      "step": 500
+    },
+    {
+      "epoch": 0.04,
+      "learning_rate": 6.991933809296747e-05,
+      "loss": 0.5263,
+      "step": 600
+    },
+    {
+      "epoch": 0.05,
+      "learning_rate": 6.988963919520753e-05,
+      "loss": 0.5068,
+      "step": 700
+    },
+    {
+      "epoch": 0.06,
+      "learning_rate": 6.985530306997431e-05,
+      "loss": 0.4704,
+      "step": 800
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 6.981633428092705e-05,
+      "loss": 0.4521,
+      "step": 900
+    },
+    {
+      "epoch": 0.07,
+      "learning_rate": 6.977273800745834e-05,
+      "loss": 0.4439,
+      "step": 1000
+    },
+    {
+      "epoch": 0.08,
+      "learning_rate": 6.972452004400577e-05,
+      "loss": 0.421,
+      "step": 1100
+    },
+    {
+      "epoch": 0.09,
+      "learning_rate": 6.96716867992818e-05,
+      "loss": 0.4105,
+      "step": 1200
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 6.961424529542192e-05,
+      "loss": 0.3998,
+      "step": 1300
+    },
+    {
+      "epoch": 0.1,
+      "learning_rate": 6.955220316705135e-05,
+      "loss": 0.3971,
+      "step": 1400
+    },
+    {
+      "epoch": 0.11,
+      "learning_rate": 6.948556866027035e-05,
+      "loss": 0.3958,
+      "step": 1500
+    },
+    {
+      "epoch": 0.11,
+      "eval_loss": 0.26585060358047485,
+      "eval_runtime": 1119.616,
+      "eval_samples_per_second": 14.297,
+      "eval_steps_per_second": 1.787,
+      "eval_wer": 0.2537555992570742,
+      "step": 1500
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 6.941435063155818e-05,
+      "loss": 0.3828,
+      "step": 1600
+    },
+    {
+      "epoch": 0.12,
+      "learning_rate": 6.933855854659593e-05,
+      "loss": 0.3852,
+      "step": 1700
+    },
+    {
+      "epoch": 0.13,
+      "learning_rate": 6.925820247900854e-05,
+      "loss": 0.3749,
+      "step": 1800
+    },
+    {
+      "epoch": 0.14,
+      "learning_rate": 6.917329310902582e-05,
+      "loss": 0.3853,
+      "step": 1900
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 6.90838417220629e-05,
+      "loss": 0.3692,
+      "step": 2000
+    },
+    {
+      "epoch": 0.15,
+      "learning_rate": 6.898986020722038e-05,
+      "loss": 0.3676,
+      "step": 2100
+    },
+    {
+      "epoch": 0.16,
+      "learning_rate": 6.889136105570403e-05,
+      "loss": 0.3697,
+      "step": 2200
+    },
+    {
+      "epoch": 0.17,
+      "learning_rate": 6.878835735916458e-05,
+      "loss": 0.3608,
+      "step": 2300
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 6.868086280795778e-05,
+      "loss": 0.3515,
+      "step": 2400
+    },
+    {
+      "epoch": 0.18,
+      "learning_rate": 6.85688916893247e-05,
+      "loss": 0.3383,
+      "step": 2500
+    },
+    {
+      "epoch": 0.19,
+      "learning_rate": 6.845245888549281e-05,
+      "loss": 0.3421,
+      "step": 2600
+    },
+    {
+      "epoch": 0.2,
+      "learning_rate": 6.833157987169802e-05,
+      "loss": 0.3454,
+      "step": 2700
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 6.820627071412778e-05,
+      "loss": 0.3499,
+      "step": 2800
+    },
+    {
+      "epoch": 0.21,
+      "learning_rate": 6.807654806778575e-05,
+      "loss": 0.3411,
+      "step": 2900
+    },
+    {
+      "epoch": 0.22,
+      "learning_rate": 6.794242917427811e-05,
+      "loss": 0.3422,
+      "step": 3000
+    },
+    {
+      "epoch": 0.22,
+      "eval_loss": 0.21751175820827484,
+      "eval_runtime": 1050.6432,
+      "eval_samples_per_second": 15.235,
+      "eval_steps_per_second": 1.905,
+      "eval_wer": 0.23247159401289194,
+      "step": 3000
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 6.780393185952203e-05,
+      "loss": 0.3378,
+      "step": 3100
+    },
+    {
+      "epoch": 0.23,
+      "learning_rate": 6.766107453137634e-05,
+      "loss": 0.3354,
+      "step": 3200
+    },
+    {
+      "epoch": 0.24,
+      "learning_rate": 6.751387617719493e-05,
+      "loss": 0.3332,
+      "step": 3300
+    },
+    {
+      "epoch": 0.25,
+      "learning_rate": 6.736235636130315e-05,
+      "loss": 0.3313,
+      "step": 3400
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 6.720653522239741e-05,
+      "loss": 0.3291,
+      "step": 3500
+    },
+    {
+      "epoch": 0.26,
+      "learning_rate": 6.704643347086866e-05,
+      "loss": 0.3225,
+      "step": 3600
+    },
+    {
+      "epoch": 0.27,
+      "learning_rate": 6.688207238604962e-05,
+      "loss": 0.3314,
+      "step": 3700
+    },
+    {
+      "epoch": 0.28,
+      "learning_rate": 6.671347381338648e-05,
+      "loss": 0.3279,
+      "step": 3800
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 6.654066016153562e-05,
+      "loss": 0.3248,
+      "step": 3900
+    },
+    {
+      "epoch": 0.29,
+      "learning_rate": 6.636365439938497e-05,
+      "loss": 0.3184,
+      "step": 4000
+    },
+    {
+      "epoch": 0.3,
+      "learning_rate": 6.618248005300135e-05,
+      "loss": 0.3312,
+      "step": 4100
+    },
+    {
+      "epoch": 0.31,
+      "learning_rate": 6.599716120250359e-05,
+      "loss": 0.3156,
+      "step": 4200
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 6.58077224788619e-05,
+      "loss": 0.32,
+      "step": 4300
+    },
+    {
+      "epoch": 0.32,
+      "learning_rate": 6.561418906062424e-05,
+      "loss": 0.3143,
+      "step": 4400
+    },
+    {
+      "epoch": 0.33,
+      "learning_rate": 6.541658667056979e-05,
+      "loss": 0.3151,
+      "step": 4500
+    },
+    {
+      "epoch": 0.33,
+      "eval_loss": 0.2139398157596588,
+      "eval_runtime": 1046.9527,
+      "eval_samples_per_second": 15.289,
+      "eval_steps_per_second": 1.911,
+      "eval_wer": 0.19961351469463565,
+      "step": 4500
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 6.521494157229007e-05,
+      "loss": 0.3228,
+      "step": 4600
+    },
+    {
+      "epoch": 0.34,
+      "learning_rate": 6.50092805666982e-05,
+      "loss": 0.3223,
+      "step": 4700
+    },
+    {
+      "epoch": 0.35,
+      "learning_rate": 6.47996309884668e-05,
+      "loss": 0.3186,
+      "step": 4800
+    },
+    {
+      "epoch": 0.36,
+      "learning_rate": 6.45860207023949e-05,
+      "loss": 0.3069,
+      "step": 4900
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 6.436847809970438e-05,
+      "loss": 0.3149,
+      "step": 5000
+    },
+    {
+      "epoch": 0.37,
+      "learning_rate": 6.41470320942664e-05,
+      "loss": 0.3152,
+      "step": 5100
+    },
+    {
+      "epoch": 0.38,
+      "learning_rate": 6.392171211875852e-05,
+      "loss": 0.3054,
+      "step": 5200
+    },
+    {
+      "epoch": 0.39,
+      "learning_rate": 6.369485868905532e-05,
+      "loss": 0.3196,
+      "step": 5300
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 6.346191911024053e-05,
+      "loss": 0.3011,
+      "step": 5400
+    },
+    {
+      "epoch": 0.4,
+      "learning_rate": 6.322519662061658e-05,
+      "loss": 0.3103,
+      "step": 5500
+    },
+    {
+      "epoch": 0.41,
+      "learning_rate": 6.298472268327846e-05,
+      "loss": 0.305,
+      "step": 5600
+    },
+    {
+      "epoch": 0.42,
+      "learning_rate": 6.274052925993097e-05,
+      "loss": 0.3043,
+      "step": 5700
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 6.249264880664065e-05,
+      "loss": 0.3031,
+      "step": 5800
+    },
+    {
+      "epoch": 0.43,
+      "learning_rate": 6.224111426952202e-05,
+      "loss": 0.2988,
+      "step": 5900
+    },
+    {
+      "epoch": 0.44,
+      "learning_rate": 6.198595908035864e-05,
+      "loss": 0.3,
+      "step": 6000
+    },
+    {
+      "epoch": 0.44,
+      "eval_loss": 0.20627757906913757,
+      "eval_runtime": 1055.8743,
+      "eval_samples_per_second": 15.16,
+      "eval_steps_per_second": 1.895,
+      "eval_wer": 0.20541079427510106,
+      "step": 6000
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 6.172721715215964e-05,
+      "loss": 0.2983,
+      "step": 6100
+    },
+    {
+      "epoch": 0.45,
+      "learning_rate": 6.146492287465236e-05,
+      "loss": 0.3012,
+      "step": 6200
+    },
+    {
+      "epoch": 0.46,
+      "learning_rate": 6.119911110971146e-05,
+      "loss": 0.2989,
+      "step": 6300
+    },
+    {
+      "epoch": 0.47,
+      "learning_rate": 6.092981718672549e-05,
+      "loss": 0.2909,
+      "step": 6400
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 6.065707689790118e-05,
+      "loss": 0.3063,
+      "step": 6500
+    },
+    {
+      "epoch": 0.48,
+      "learning_rate": 6.03809264935062e-05,
+      "loss": 0.2908,
+      "step": 6600
+    },
+    {
+      "epoch": 0.49,
+      "learning_rate": 6.0101402677051154e-05,
+      "loss": 0.309,
+      "step": 6700
+    },
+    {
+      "epoch": 0.5,
+      "learning_rate": 5.981854260041124e-05,
+      "loss": 0.2891,
+      "step": 6800
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 5.9532383858888345e-05,
+      "loss": 0.2902,
+      "step": 6900
+    },
+    {
+      "epoch": 0.51,
+      "learning_rate": 5.924296448621422e-05,
+      "loss": 0.289,
+      "step": 7000
+    },
+    {
+      "epoch": 0.52,
+      "learning_rate": 5.8950322949495356e-05,
+      "loss": 0.2951,
+      "step": 7100
+    },
+    {
+      "epoch": 0.53,
+      "learning_rate": 5.8654498144100274e-05,
+      "loss": 0.2923,
+      "step": 7200
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 5.835552938848987e-05,
+      "loss": 0.2967,
+      "step": 7300
+    },
+    {
+      "epoch": 0.54,
+      "learning_rate": 5.805345641899159e-05,
+      "loss": 0.2863,
+      "step": 7400
+    },
+    {
+      "epoch": 0.55,
+      "learning_rate": 5.774831938451798e-05,
+      "loss": 0.2824,
+      "step": 7500
+    },
+    {
+      "epoch": 0.55,
+      "eval_loss": 0.18351121246814728,
+      "eval_runtime": 1075.697,
+      "eval_samples_per_second": 14.881,
+      "eval_steps_per_second": 1.86,
+      "eval_wer": 0.18133398885611274,
+      "step": 7500
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 5.744325527887681e-05,
+      "loss": 0.2833,
+      "step": 7600
+    },
+    {
+      "epoch": 0.56,
+      "learning_rate": 5.713214180625491e-05,
+      "loss": 0.2909,
+      "step": 7700
+    },
+    {
+      "epoch": 0.57,
+      "learning_rate": 5.68180867217851e-05,
+      "loss": 0.286,
+      "step": 7800
+    },
+    {
+      "epoch": 0.58,
+      "learning_rate": 5.650113176693846e-05,
+      "loss": 0.2778,
+      "step": 7900
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 5.618131906861165e-05,
+      "loss": 0.2777,
+      "step": 8000
+    },
+    {
+      "epoch": 0.59,
+      "learning_rate": 5.5858691133527713e-05,
+      "loss": 0.2829,
+      "step": 8100
+    },
+    {
+      "epoch": 0.6,
+      "learning_rate": 5.553329084258652e-05,
+      "loss": 0.2748,
+      "step": 8200
+    },
+    {
+      "epoch": 0.61,
+      "learning_rate": 5.5205161445165346e-05,
+      "loss": 0.2827,
+      "step": 8300
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 5.4874346553370585e-05,
+      "loss": 0.2712,
+      "step": 8400
+    },
+    {
+      "epoch": 0.62,
+      "learning_rate": 5.4540890136241195e-05,
+      "loss": 0.2792,
+      "step": 8500
+    },
+    {
+      "epoch": 0.63,
+      "learning_rate": 5.420483651390469e-05,
+      "loss": 0.2815,
+      "step": 8600
+    },
+    {
+      "epoch": 0.64,
+      "learning_rate": 5.386623035168656e-05,
+      "loss": 0.2766,
+      "step": 8700
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 5.3525116654173646e-05,
+      "loss": 0.2704,
+      "step": 8800
+    },
+    {
+      "epoch": 0.65,
+      "learning_rate": 5.318154075923263e-05,
+      "loss": 0.2695,
+      "step": 8900
+    },
+    {
+      "epoch": 0.66,
+      "learning_rate": 5.283554833198404e-05,
+      "loss": 0.2771,
+      "step": 9000
+    },
+    {
+      "epoch": 0.66,
+      "eval_loss": 0.18808312714099884,
+      "eval_runtime": 1061.3581,
+      "eval_samples_per_second": 15.082,
+      "eval_steps_per_second": 1.885,
+      "eval_wer": 0.17624685895334863,
+      "step": 9000
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 5.2487185358732866e-05,
+      "loss": 0.276,
+      "step": 9100
+    },
+    {
+      "epoch": 0.67,
+      "learning_rate": 5.213649814085646e-05,
+      "loss": 0.2647,
+      "step": 9200
+    },
+    {
+      "epoch": 0.68,
+      "learning_rate": 5.178353328865057e-05,
+      "loss": 0.2901,
+      "step": 9300
+    },
+    {
+      "epoch": 0.69,
+      "learning_rate": 5.142833771513431e-05,
+      "loss": 0.2771,
+      "step": 9400
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 5.107095862981481e-05,
+      "loss": 0.2706,
+      "step": 9500
+    },
+    {
+      "epoch": 0.7,
+      "learning_rate": 5.071144353241269e-05,
+      "loss": 0.2753,
+      "step": 9600
+    },
+    {
+      "epoch": 0.71,
+      "learning_rate": 5.0353466418956284e-05,
+      "loss": 0.2777,
+      "step": 9700
+    },
+    {
+      "epoch": 0.72,
+      "learning_rate": 4.9989843088674705e-05,
+      "loss": 0.2621,
+      "step": 9800
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 4.962422743878782e-05,
+      "loss": 0.2628,
+      "step": 9900
+    },
+    {
+      "epoch": 0.73,
+      "learning_rate": 4.9256668063748734e-05,
+      "loss": 0.272,
+      "step": 10000
+    },
+    {
+      "epoch": 0.74,
+      "learning_rate": 4.8887213816353655e-05,
+      "loss": 0.2683,
+      "step": 10100
+    },
+    {
+      "epoch": 0.75,
+      "learning_rate": 4.851591380124868e-05,
+      "loss": 0.2627,
+      "step": 10200
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 4.814281736840332e-05,
+      "loss": 0.2565,
+      "step": 10300
+    },
+    {
+      "epoch": 0.76,
+      "learning_rate": 4.776797410655135e-05,
+      "loss": 0.2619,
+      "step": 10400
+    },
+    {
+      "epoch": 0.77,
+      "learning_rate": 4.739143383659982e-05,
+      "loss": 0.2616,
+      "step": 10500
+    },
+    {
+      "epoch": 0.77,
+      "eval_loss": 0.18062810599803925,
+      "eval_runtime": 1044.5923,
+      "eval_samples_per_second": 15.324,
+      "eval_steps_per_second": 1.916,
+      "eval_wer": 0.17676581448705342,
+      "step": 10500
+    },
+    {
+      "epoch": 0.78,
+      "learning_rate": 4.701324660500736e-05,
+      "loss": 0.2575,
+      "step": 10600
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 4.663346267713244e-05,
+      "loss": 0.2642,
+      "step": 10700
+    },
+    {
+      "epoch": 0.79,
+      "learning_rate": 4.625213253055248e-05,
+      "loss": 0.2561,
+      "step": 10800
+    },
+    {
+      "epoch": 0.8,
+      "learning_rate": 4.586930684835486e-05,
+      "loss": 0.2553,
+      "step": 10900
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 4.5485036512400575e-05,
+      "loss": 0.2533,
+      "step": 11000
+    },
+    {
+      "epoch": 0.81,
+      "learning_rate": 4.509937259656139e-05,
+      "loss": 0.2528,
+      "step": 11100
+    },
+    {
+      "epoch": 0.82,
+      "learning_rate": 4.471236635993164e-05,
+      "loss": 0.2542,
+      "step": 11200
+    },
+    {
+      "epoch": 0.83,
+      "learning_rate": 4.432406924001522e-05,
+      "loss": 0.2506,
+      "step": 11300
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 4.393453284588905e-05,
+      "loss": 0.2587,
+      "step": 11400
+    },
+    {
+      "epoch": 0.84,
+      "learning_rate": 4.3543808951343574e-05,
+      "loss": 0.2503,
+      "step": 11500
+    },
+    {
+      "epoch": 0.85,
+      "learning_rate": 4.3151949488001475e-05,
+      "loss": 0.2535,
+      "step": 11600
+    },
+    {
+      "epoch": 0.86,
+      "learning_rate": 4.275900653841536e-05,
+      "loss": 0.2526,
+      "step": 11700
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 4.236503232914543e-05,
+      "loss": 0.253,
+      "step": 11800
+    },
+    {
+      "epoch": 0.87,
+      "learning_rate": 4.197007922381793e-05,
+      "loss": 0.2523,
+      "step": 11900
+    },
+    {
+      "epoch": 0.88,
+      "learning_rate": 4.157419971616547e-05,
+      "loss": 0.2446,
+      "step": 12000
+    },
+    {
+      "epoch": 0.88,
+      "eval_loss": 0.1757470816373825,
+      "eval_runtime": 1062.3089,
+      "eval_samples_per_second": 15.068,
+      "eval_steps_per_second": 1.884,
+      "eval_wer": 0.1589711023708074,
+      "step": 12000
+    },
+    {
+      "epoch": 0.89,
+      "learning_rate": 4.1177446423050005e-05,
+      "loss": 0.2473,
+      "step": 12100
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 4.077987207746943e-05,
+      "loss": 0.2438,
+      "step": 12200
+    },
+    {
+      "epoch": 0.9,
+      "learning_rate": 4.0381529521548834e-05,
+      "loss": 0.2468,
+      "step": 12300
+    },
+    {
+      "epoch": 0.91,
+      "learning_rate": 3.998247169951711e-05,
+      "loss": 0.2524,
+      "step": 12400
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 3.958275165067014e-05,
+      "loss": 0.2372,
+      "step": 12500
+    },
+    {
+      "epoch": 0.92,
+      "learning_rate": 3.91824225023212e-05,
+      "loss": 0.2324,
+      "step": 12600
+    },
+    {
+      "epoch": 0.93,
+      "learning_rate": 3.8785548889903e-05,
+      "loss": 0.2436,
+      "step": 12700
+    },
+    {
+      "epoch": 0.94,
+      "learning_rate": 3.8384166003361756e-05,
+      "loss": 0.2405,
+      "step": 12800
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 3.798635370086602e-05,
+      "loss": 0.2349,
+      "step": 12900
+    },
+    {
+      "epoch": 0.95,
+      "learning_rate": 3.7584128333900755e-05,
+      "loss": 0.2424,
+      "step": 13000
+    },
+    {
+      "epoch": 0.96,
+      "learning_rate": 3.7181559507066575e-05,
+      "loss": 0.2366,
+      "step": 13100
+    },
+    {
+      "epoch": 0.97,
+      "learning_rate": 3.677870072631157e-05,
+      "loss": 0.2435,
+      "step": 13200
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 3.637560553612199e-05,
+      "loss": 0.2377,
+      "step": 13300
+    },
+    {
+      "epoch": 0.98,
+      "learning_rate": 3.597232751240556e-05,
+      "loss": 0.2302,
+      "step": 13400
+    },
+    {
+      "epoch": 0.99,
+      "learning_rate": 3.556892025537066e-05,
+      "loss": 0.2377,
+      "step": 13500
+    },
+    {
+      "epoch": 0.99,
+      "eval_loss": 0.1588028222322464,
+      "eval_runtime": 1044.908,
+      "eval_samples_per_second": 15.319,
+      "eval_steps_per_second": 1.915,
+      "eval_wer": 0.15275046432863543,
+      "step": 13500
+    },
+    {
+      "epoch": 1.0,
+      "learning_rate": 3.516543738240223e-05,
+      "loss": 0.2313,
+      "step": 13600
+    },
+    {
+      "epoch": 1.01,
+      "learning_rate": 3.476193252093543e-05,
+      "loss": 0.22,
+      "step": 13700
+    },
+    {
+      "epoch": 1.01,
+      "learning_rate": 3.4358459301327927e-05,
+      "loss": 0.214,
+      "step": 13800
+    },
+    {
+      "epoch": 1.02,
+      "learning_rate": 3.395507134973183e-05,
+      "loss": 0.2257,
+      "step": 13900
+    },
+    {
+      "epoch": 1.03,
+      "learning_rate": 3.355182228096618e-05,
+      "loss": 0.2308,
+      "step": 14000
+    },
+    {
+      "epoch": 1.03,
+      "learning_rate": 3.314876569139091e-05,
+      "loss": 0.2244,
+      "step": 14100
+    },
+    {
+      "epoch": 1.04,
+      "learning_rate": 3.274595515178329e-05,
+      "loss": 0.2176,
+      "step": 14200
+    },
+    {
+      "epoch": 1.05,
+      "learning_rate": 3.234344420021777e-05,
+      "loss": 0.2238,
+      "step": 14300
+    },
+    {
+      "epoch": 1.06,
+      "learning_rate": 3.194128633495017e-05,
+      "loss": 0.219,
+      "step": 14400
+    },
+    {
+      "epoch": 1.06,
+      "learning_rate": 3.153953500730713e-05,
+      "loss": 0.2265,
+      "step": 14500
+    },
+    {
+      "epoch": 1.07,
+      "learning_rate": 3.113824361458186e-05,
+      "loss": 0.2218,
+      "step": 14600
+    },
+    {
+      "epoch": 1.08,
+      "learning_rate": 3.073746549293703e-05,
+      "loss": 0.2129,
+      "step": 14700
+    },
+    {
+      "epoch": 1.09,
+      "learning_rate": 3.0337253910315748e-05,
+      "loss": 0.2126,
+      "step": 14800
+    },
+    {
+      "epoch": 1.09,
+      "learning_rate": 2.993766205936171e-05,
+      "loss": 0.2047,
+      "step": 14900
+    },
+    {
+      "epoch": 1.1,
+      "learning_rate": 2.9538743050349254e-05,
+      "loss": 0.2141,
+      "step": 15000
+    },
+    {
+      "epoch": 1.1,
+      "eval_loss": 0.14499780535697937,
+      "eval_runtime": 1046.3411,
+      "eval_samples_per_second": 15.298,
+      "eval_steps_per_second": 1.912,
+      "eval_wer": 0.14962307440183548,
+      "step": 15000
+    },
+    {
+      "epoch": 1.11,
+      "learning_rate": 2.9140549904124422e-05,
+      "loss": 0.2066,
+      "step": 15100
+    },
+    {
+      "epoch": 1.12,
+      "learning_rate": 2.8743135545057887e-05,
+      "loss": 0.2124,
+      "step": 15200
+    },
+    {
+      "epoch": 1.12,
+      "learning_rate": 2.8346552794010703e-05,
+      "loss": 0.2089,
+      "step": 15300
+    },
+    {
+      "epoch": 1.13,
+      "learning_rate": 2.7950854361313814e-05,
+      "loss": 0.2121,
+      "step": 15400
+    },
+    {
+      "epoch": 1.14,
+      "learning_rate": 2.755609283976226e-05,
+      "loss": 0.209,
+      "step": 15500
+    },
+    {
+      "epoch": 1.14,
+      "learning_rate": 2.7162320697625e-05,
+      "loss": 0.2052,
+      "step": 15600
+    },
+    {
+      "epoch": 1.15,
+      "learning_rate": 2.676959027167128e-05,
+      "loss": 0.209,
+      "step": 15700
+    },
+    {
+      "epoch": 1.16,
+      "learning_rate": 2.6377953760214495e-05,
+      "loss": 0.2089,
+      "step": 15800
+    },
+    {
+      "epoch": 1.17,
+      "learning_rate": 2.598746321617443e-05,
+      "loss": 0.2021,
+      "step": 15900
+    },
+    {
+      "epoch": 1.17,
+      "learning_rate": 2.5598170540158846e-05,
+      "loss": 0.2047,
+      "step": 16000
+    },
+    {
+      "epoch": 1.18,
+      "learning_rate": 2.5210127473565314e-05,
+      "loss": 0.1979,
+      "step": 16100
+    },
+    {
+      "epoch": 1.19,
+      "learning_rate": 2.482338559170417e-05,
+      "loss": 0.2044,
+      "step": 16200
+    },
+    {
+      "epoch": 1.2,
+      "learning_rate": 2.4437996296943596e-05,
+      "loss": 0.2082,
+      "step": 16300
+    },
+    {
+      "epoch": 1.2,
+      "learning_rate": 2.4057843550135512e-05,
+      "loss": 0.2004,
+      "step": 16400
+    },
+    {
+      "epoch": 1.21,
+      "learning_rate": 2.3675298110320073e-05,
+      "loss": 0.1953,
+      "step": 16500
+    },
+    {
+      "epoch": 1.21,
+      "eval_loss": 0.13918258249759674,
+      "eval_runtime": 1060.8068,
+      "eval_samples_per_second": 15.089,
+      "eval_steps_per_second": 1.886,
+      "eval_wer": 0.13435485633125752,
+      "step": 16500
+    },
+    {
+      "epoch": 1.22,
+      "learning_rate": 2.3294257851410495e-05,
+      "loss": 0.1984,
+      "step": 16600
+    },
+    {
+      "epoch": 1.23,
+      "learning_rate": 2.2914773417964826e-05,
+      "loss": 0.1972,
+      "step": 16700
+    },
+    {
+      "epoch": 1.23,
+      "learning_rate": 2.2536895247754305e-05,
+      "loss": 0.194,
+      "step": 16800
+    },
+    {
+      "epoch": 1.24,
+      "learning_rate": 2.2160673565059625e-05,
+      "loss": 0.2016,
+      "step": 16900
+    },
+    {
+      "epoch": 1.25,
+      "learning_rate": 2.1786158373995577e-05,
+      "loss": 0.1973,
+      "step": 17000
+    },
+    {
+      "epoch": 1.25,
+      "learning_rate": 2.1413399451864916e-05,
+      "loss": 0.1872,
+      "step": 17100
+    },
+    {
+      "epoch": 1.26,
+      "learning_rate": 2.1042446342542387e-05,
+      "loss": 0.2035,
+      "step": 17200
+    },
+    {
+      "epoch": 1.27,
+      "learning_rate": 2.0673348349889817e-05,
+      "loss": 0.1937,
+      "step": 17300
+    },
+    {
+      "epoch": 1.28,
+      "learning_rate": 2.0306154531203048e-05,
+      "loss": 0.1938,
+      "step": 17400
+    },
+    {
+      "epoch": 1.28,
+      "learning_rate": 1.994091369069168e-05,
+      "loss": 0.1953,
+      "step": 17500
+    },
+    {
+      "epoch": 1.29,
+      "learning_rate": 1.957767437299243e-05,
+      "loss": 0.1887,
+      "step": 17600
+    },
+    {
+      "epoch": 1.3,
+      "learning_rate": 1.9216484856717008e-05,
+      "loss": 0.1921,
+      "step": 17700
+    },
+    {
+      "epoch": 1.31,
+      "learning_rate": 1.8857393148035336e-05,
+      "loss": 0.1917,
+      "step": 17800
+    },
+    {
+      "epoch": 1.31,
+      "learning_rate": 1.8500446974295e-05,
+      "loss": 0.1836,
+      "step": 17900
+    },
+    {
+      "epoch": 1.32,
+      "learning_rate": 1.8145693777677743e-05,
+      "loss": 0.1923,
+      "step": 18000
+    },
+    {
+      "epoch": 1.32,
+      "eval_loss": 0.1327279508113861,
+      "eval_runtime": 1039.577,
+      "eval_samples_per_second": 15.398,
+      "eval_steps_per_second": 1.925,
+      "eval_wer": 0.13173959357587675,
+      "step": 18000
+    },
+    {
+      "epoch": 1.33,
+      "learning_rate": 1.779669459685722e-05,
+      "loss": 0.1927,
+      "step": 18100
+    },
+    {
+      "epoch": 1.34,
+      "learning_rate": 1.7446445408141307e-05,
+      "loss": 0.1959,
+      "step": 18200
+    },
+    {
+      "epoch": 1.34,
+      "learning_rate": 1.7098529285272e-05,
+      "loss": 0.1872,
+      "step": 18300
+    },
+    {
+      "epoch": 1.35,
+      "learning_rate": 1.6752992470235188e-05,
+      "loss": 0.1848,
+      "step": 18400
+    },
+    {
+      "epoch": 1.36,
+      "learning_rate": 1.640988088877985e-05,
+      "loss": 0.184,
+      "step": 18500
+    },
+    {
+      "epoch": 1.36,
+      "learning_rate": 1.6069240144314012e-05,
+      "loss": 0.1919,
+      "step": 18600
+    },
+    {
+      "epoch": 1.37,
+      "learning_rate": 1.5731115511843525e-05,
+      "loss": 0.176,
+      "step": 18700
+    },
+    {
+      "epoch": 1.38,
+      "learning_rate": 1.5395551931954524e-05,
+      "loss": 0.1789,
+      "step": 18800
+    },
+    {
+      "epoch": 1.39,
+      "learning_rate": 1.5062594004840269e-05,
+      "loss": 0.1937,
+      "step": 18900
+    },
+    {
+      "epoch": 1.39,
+      "learning_rate": 1.4732285984373345e-05,
+      "loss": 0.1861,
+      "step": 19000
+    },
+    {
+      "epoch": 1.4,
+      "learning_rate": 1.440467177222377e-05,
+      "loss": 0.1757,
+      "step": 19100
+    },
+    {
+      "epoch": 1.41,
+      "learning_rate": 1.4079794912023988e-05,
+      "loss": 0.1839,
+      "step": 19200
+    },
+    {
+      "epoch": 1.42,
+      "learning_rate": 1.3757698583581431e-05,
+      "loss": 0.1844,
+      "step": 19300
+    },
+    {
+      "epoch": 1.42,
+      "learning_rate": 1.3438425597139414e-05,
+      "loss": 0.1855,
+      "step": 19400
+    },
+    {
+      "epoch": 1.43,
+      "learning_rate": 1.3122018387687183e-05,
+      "loss": 0.1804,
+      "step": 19500
+    },
+    {
+      "epoch": 1.43,
+      "eval_loss": 0.12711018323898315,
+      "eval_runtime": 1058.0447,
+      "eval_samples_per_second": 15.129,
+      "eval_steps_per_second": 1.891,
+      "eval_wer": 0.12372992461488037,
+      "step": 19500
+    },
+    {
+      "epoch": 1.44,
+      "learning_rate": 1.280851900931984e-05,
+      "loss": 0.1833,
+      "step": 19600
+    },
+    {
+      "epoch": 1.45,
+      "learning_rate": 1.2497969129648841e-05,
+      "loss": 0.181,
+      "step": 19700
+    },
+    {
+      "epoch": 1.45,
+      "learning_rate": 1.2190410024263938e-05,
+      "loss": 0.1719,
+      "step": 19800
+    },
+    {
+      "epoch": 1.46,
+      "learning_rate": 1.1885882571247166e-05,
+      "loss": 0.1758,
+      "step": 19900
+    },
+    {
+      "epoch": 1.47,
+      "learning_rate": 1.1584427245739682e-05,
+      "loss": 0.1792,
+      "step": 20000
+    },
+    {
+      "epoch": 1.47,
+      "learning_rate": 1.1286084114562175e-05,
+      "loss": 0.1774,
+      "step": 20100
+    },
+    {
+      "epoch": 1.48,
+      "learning_rate": 1.0990892830889517e-05,
+      "loss": 0.1796,
+      "step": 20200
+    },
+    {
+      "epoch": 1.49,
+      "learning_rate": 1.0698892628980422e-05,
+      "loss": 0.1816,
+      "step": 20300
+    },
+    {
+      "epoch": 1.5,
+      "learning_rate": 1.041012231896276e-05,
+      "loss": 0.174,
+      "step": 20400
+    },
+    {
+      "epoch": 1.5,
+      "learning_rate": 1.012462028167525e-05,
+      "loss": 0.1717,
+      "step": 20500
+    },
+    {
+      "epoch": 1.51,
+      "learning_rate": 9.842424463566227e-06,
+      "loss": 0.1793,
+      "step": 20600
+    },
+    {
+      "epoch": 1.52,
+      "learning_rate": 9.563572371650113e-06,
+      "loss": 0.1699,
+      "step": 20700
+    },
+    {
+      "epoch": 1.53,
+      "learning_rate": 9.288101068522322e-06,
+      "loss": 0.1726,
+      "step": 20800
+    },
+    {
+      "epoch": 1.53,
+      "learning_rate": 9.016047167433221e-06,
+      "loss": 0.1734,
+      "step": 20900
+    },
+    {
+      "epoch": 1.54,
+      "learning_rate": 8.747446827421805e-06,
+      "loss": 0.1776,
+      "step": 21000
+    },
+    {
+      "epoch": 1.54,
+      "eval_loss": 0.12307832390069962,
+      "eval_runtime": 1040.1826,
+      "eval_samples_per_second": 15.389,
+      "eval_steps_per_second": 1.924,
+      "eval_wer": 0.1186018245384027,
+      "step": 21000
+    },
+    {
+      "epoch": 1.55,
+      "learning_rate": 8.482335748509769e-06,
+      "loss": 0.1755,
+      "step": 21100
+    },
+    {
+      "epoch": 1.56,
+      "learning_rate": 8.220749166956552e-06,
+      "loss": 0.1717,
+      "step": 21200
+    },
+    {
+      "epoch": 1.56,
+      "learning_rate": 7.962721850576054e-06,
+      "loss": 0.167,
+      "step": 21300
+    },
+    {
+      "epoch": 1.57,
+      "learning_rate": 7.708288094115607e-06,
+      "loss": 0.1698,
+      "step": 21400
+    },
+    {
+      "epoch": 1.58,
+      "learning_rate": 7.457481714697784e-06,
+      "loss": 0.1709,
+      "step": 21500
+    },
+    {
+      "epoch": 1.58,
+      "learning_rate": 7.210336047325761e-06,
+      "loss": 0.1748,
+      "step": 21600
+    },
+    {
+      "epoch": 1.59,
+      "learning_rate": 6.9668839404526865e-06,
+      "loss": 0.1776,
+      "step": 21700
+    },
+    {
+      "epoch": 1.6,
+      "learning_rate": 6.727157751615771e-06,
+      "loss": 0.1664,
+      "step": 21800
+    },
+    {
+      "epoch": 1.61,
+      "learning_rate": 6.491189343135589e-06,
+      "loss": 0.1754,
+      "step": 21900
+    },
+    {
+      "epoch": 1.61,
+      "learning_rate": 6.2590100778812376e-06,
+      "loss": 0.1766,
+      "step": 22000
+    },
+    {
+      "epoch": 1.62,
+      "learning_rate": 6.030650815101828e-06,
+      "loss": 0.171,
+      "step": 22100
+    },
+    {
+      "epoch": 1.63,
+      "learning_rate": 5.808367837755271e-06,
+      "loss": 0.1703,
+      "step": 22200
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 5.5877001747984834e-06,
+      "loss": 0.164,
+      "step": 22300
+    },
+    {
+      "epoch": 1.64,
+      "learning_rate": 5.3709417389918604e-06,
+      "loss": 0.1664,
+      "step": 22400
+    },
+    {
+      "epoch": 1.65,
+      "learning_rate": 5.158121339981953e-06,
+      "loss": 0.1671,
+      "step": 22500
+    },
+    {
+      "epoch": 1.65,
+      "eval_loss": 0.11993325501680374,
+      "eval_runtime": 1045.6599,
+      "eval_samples_per_second": 15.308,
+      "eval_steps_per_second": 1.914,
+      "eval_wer": 0.11479159838304381,
+      "step": 22500
+    },
+    {
+      "epoch": 1.66,
+      "learning_rate": 4.949267264005701e-06,
+      "loss": 0.1576,
+      "step": 22600
+    },
+    {
+      "epoch": 1.67,
+      "learning_rate": 4.7444072701308795e-06,
+      "loss": 0.1583,
+      "step": 22700
+    },
+    {
+      "epoch": 1.67,
+      "learning_rate": 4.543568586566601e-06,
+      "loss": 0.1678,
+      "step": 22800
+    },
+    {
+      "epoch": 1.68,
+      "learning_rate": 4.346777907044375e-06,
+      "loss": 0.1687,
+      "step": 22900
+    },
+    {
+      "epoch": 1.69,
+      "learning_rate": 4.154061387270205e-06,
+      "loss": 0.1671,
+      "step": 23000
+    },
+    {
+      "epoch": 1.69,
+      "learning_rate": 3.965444641448219e-06,
+      "loss": 0.1656,
+      "step": 23100
+    },
+    {
+      "epoch": 1.7,
+      "learning_rate": 3.780952738876231e-06,
+      "loss": 0.169,
+      "step": 23200
+    },
+    {
+      "epoch": 1.71,
+      "learning_rate": 3.600610200613753e-06,
+      "loss": 0.1619,
+      "step": 23300
+    },
+    {
+      "epoch": 1.72,
+      "learning_rate": 3.4244409962228724e-06,
+      "loss": 0.1702,
+      "step": 23400
+    },
+    {
+      "epoch": 1.72,
+      "learning_rate": 3.252468540582438e-06,
+      "loss": 0.1654,
+      "step": 23500
+    },
+    {
+      "epoch": 1.73,
+      "learning_rate": 3.0847156907759337e-06,
+      "loss": 0.1593,
+      "step": 23600
+    },
+    {
+      "epoch": 1.74,
+      "learning_rate": 2.92120474305353e-06,
+      "loss": 0.1622,
+      "step": 23700
+    },
+    {
+      "epoch": 1.75,
+      "learning_rate": 2.7619574298686577e-06,
+      "loss": 0.1653,
+      "step": 23800
+    },
+    {
+      "epoch": 1.75,
+      "learning_rate": 2.6069949169895127e-06,
+      "loss": 0.1637,
+      "step": 23900
+    },
+    {
+      "epoch": 1.76,
+      "learning_rate": 2.4578229939112028e-06,
+      "loss": 0.1597,
+      "step": 24000
+    },
+    {
+      "epoch": 1.76,
+      "eval_loss": 0.11753135174512863,
+      "eval_runtime": 1040.817,
+      "eval_samples_per_second": 15.379,
+      "eval_steps_per_second": 1.923,
+      "eval_wer": 0.11268846279908226,
+      "step": 24000
+    },
+    {
+      "epoch": 1.77,
+      "learning_rate": 2.311447946777479e-06,
+      "loss": 0.1599,
+      "step": 24100
+    },
+    {
+      "epoch": 1.78,
+      "learning_rate": 2.1694175777527574e-06,
+      "loss": 0.1628,
+      "step": 24200
+    },
+    {
+      "epoch": 1.78,
+      "learning_rate": 2.0317507642787156e-06,
+      "loss": 0.1638,
+      "step": 24300
+    },
+    {
+      "epoch": 1.79,
+      "learning_rate": 1.898465803831184e-06,
+      "loss": 0.1651,
+      "step": 24400
+    },
+    {
+      "epoch": 1.8,
+      "learning_rate": 1.7695804114881745e-06,
+      "loss": 0.1629,
+      "step": 24500
+    },
+    {
+      "epoch": 1.8,
+      "learning_rate": 1.6451117175753708e-06,
+      "loss": 0.1699,
+      "step": 24600
+    },
+    {
+      "epoch": 1.81,
+      "learning_rate": 1.5250762653892972e-06,
+      "loss": 0.1578,
+      "step": 24700
+    },
+    {
+      "epoch": 1.82,
+      "learning_rate": 1.4094900089985423e-06,
+      "loss": 0.1648,
+      "step": 24800
+    },
+    {
+      "epoch": 1.83,
+      "learning_rate": 1.2983683111232683e-06,
+      "loss": 0.1607,
+      "step": 24900
+    },
+    {
+      "epoch": 1.83,
+      "learning_rate": 1.1917259410933516e-06,
+      "loss": 0.1593,
+      "step": 25000
+    },
+    {
+      "epoch": 1.84,
+      "learning_rate": 1.0895770728853425e-06,
+      "loss": 0.159,
+      "step": 25100
+    },
+    {
+      "epoch": 1.85,
+      "learning_rate": 9.919352832386174e-07,
+      "loss": 0.1608,
+      "step": 25200
+    },
+    {
+      "epoch": 1.86,
+      "learning_rate": 8.988135498508481e-07,
+      "loss": 0.1619,
+      "step": 25300
+    },
+    {
+      "epoch": 1.86,
+      "learning_rate": 8.102242496531358e-07,
+      "loss": 0.163,
+      "step": 25400
+    },
+    {
+      "epoch": 1.87,
+      "learning_rate": 7.261791571649655e-07,
+      "loss": 0.1619,
+      "step": 25500
+    },
+    {
+      "epoch": 1.87,
+      "eval_loss": 0.11700794845819473,
+      "eval_runtime": 1036.7542,
+      "eval_samples_per_second": 15.44,
+      "eval_steps_per_second": 1.93,
+      "eval_wer": 0.11198514148366656,
+      "step": 25500
+    },
+    {
+      "epoch": 1.88,
+      "learning_rate": 6.466894429292585e-07,
+      "loss": 0.1627,
+      "step": 25600
+    },
+    {
+      "epoch": 1.89,
+      "learning_rate": 5.717656720276581e-07,
+      "loss": 0.165,
+      "step": 25700
+    },
+    {
+      "epoch": 1.89,
+      "learning_rate": 5.014178026763216e-07,
+      "loss": 0.1685,
+      "step": 25800
+    },
+    {
+      "epoch": 1.9,
+      "learning_rate": 4.356551849023648e-07,
+      "loss": 0.1632,
+      "step": 25900
+    },
+    {
+      "epoch": 1.91,
+      "learning_rate": 3.7448655930113146e-07,
+      "loss": 0.164,
+      "step": 26000
+    },
+    {
+      "epoch": 1.91,
+      "learning_rate": 3.179200558744649e-07,
+      "loss": 0.1649,
+      "step": 26100
+    },
+    {
+      "epoch": 1.92,
+      "learning_rate": 2.6596319295015436e-07,
+      "loss": 0.1583,
+      "step": 26200
+    },
+    {
+      "epoch": 1.93,
+      "learning_rate": 2.1862287618264806e-07,
+      "loss": 0.1629,
+      "step": 26300
+    },
+    {
+      "epoch": 1.94,
+      "learning_rate": 1.7630967021918575e-07,
+      "loss": 0.1608,
+      "step": 26400
+    },
+    {
+      "epoch": 1.94,
+      "learning_rate": 1.3817439628416527e-07,
+      "loss": 0.1629,
+      "step": 26500
+    },
+    {
+      "epoch": 1.95,
+      "learning_rate": 1.0467265308166828e-07,
+      "loss": 0.1656,
+      "step": 26600
+    },
+    {
+      "epoch": 1.96,
+      "learning_rate": 7.58088933720985e-08,
+      "loss": 0.1638,
+      "step": 26700
+    },
+    {
+      "epoch": 1.97,
+      "learning_rate": 5.158695347542152e-08,
+      "loss": 0.164,
+      "step": 26800
+    },
+    {
+      "epoch": 1.97,
+      "learning_rate": 3.2010052761280434e-08,
+      "loss": 0.1632,
+      "step": 26900
+    },
+    {
+      "epoch": 1.98,
+      "learning_rate": 1.708079322109368e-08,
+      "loss": 0.1664,
+      "step": 27000
+    },
+    {
+      "epoch": 1.98,
+      "eval_loss": 0.11697087436914444,
+      "eval_runtime": 1037.0691,
+      "eval_samples_per_second": 15.435,
+      "eval_steps_per_second": 1.929,
+      "eval_wer": 0.11169152190538621,
+      "step": 27000
+    }
+  ],
+  "max_steps": 27260,
+  "num_train_epochs": 2,
+  "total_flos": 4.0022174178965815e+20,
+  "trial_name": null,
+  "trial_params": null
+}

training_args.bin

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33687b9bde339fb2a61ab28ad3b77a61a2b72c0f63bd474a7eea4c3152b86aa6
+size 2991

vocab.json

@@ -0,0 +1 @@
+{"=": 1, "@": 2, "[": 3, "]": 4, "_": 5, "`": 6, "a": 7, "b": 8, "c": 9, "d": 10, "e": 11, "f": 12, "g": 13, "h": 14, "i": 15, "j": 16, "k": 17, "l": 18, "m": 19, "n": 20, "o": 21, "p": 22, "q": 23, "r": 24, "s": 25, "t": 26, "u": 27, "v": 28, "w": 29, "x": 30, "y": 31, "z": 32, "¡": 33, "§": 34, "«": 35, "°": 36, "´": 37, "µ": 38, "·": 39, "»": 40, "×": 41, "ß": 42, "à": 43, "á": 44, "â": 45, "ã": 46, "ä": 47, "å": 48, "æ": 49, "ç": 50, "è": 51, "é": 52, "ê": 53, "ë": 54, "ì": 55, "í": 56, "î": 57, "ï": 58, "ð": 59, "ñ": 60, "ò": 61, "ó": 62, "ô": 63, "õ": 64, "ö": 65, "ø": 66, "ù": 67, "ú": 68, "û": 69, "ü": 70, "ý": 71, "þ": 72, "ā": 73, "ă": 74, "ą": 75, "ć": 76, "č": 77, "ď": 78, "đ": 79, "ē": 80, "ė": 81, "ę": 82, "ě": 83, "ğ": 84, "ġ": 85, "ħ": 86, "ī": 87, "ı": 88, "ł": 89, "ń": 90, "ņ": 91, "ň": 92, "ō": 93, "ŏ": 94, "ő": 95, "œ": 96, "ř": 97, "ś": 98, "ş": 99, "š": 100, "ť": 101, "ū": 102, "ů": 103, "ź": 104, "ż": 105, "ž": 106, "ơ": 107, "ǐ": 108, "ǔ": 109, "ș": 110, "ț": 111, "ə": 112, "ʻ": 113, "ʾ": 114, "ʿ": 115, "̆": 116, "̇": 117, "̥": 118, "а": 119, "в": 120, "е": 121, "и": 122, "к": 123, "м": 124, "о": 125, "р": 126, "с": 127, "ф": 128, "ч": 129, "ш": 130, "ѹ": 131, "א": 132, "ב": 133, "נ": 134, "ע": 135, "ש": 136, "་": 137, "ན": 138, "ḫ": 139, "ṟ": 140, "ṣ": 141, "ṭ": 142, "ạ": 143, "ả": 144, "ắ": 145, "ằ": 146, "ế": 147, "ễ": 148, "ệ": 149, "ọ": 150, "ồ": 151, "ộ": 152, "ụ": 153, "ứ": 154, "‑": 155, "‚": 156, "„": 157, "‟": 158, "′": 159, "″": 160, "‹": 161, "›": 162, "→": 163, "−": 164, "≡": 165, "⟨": 166, "⟩": 167, "カ": 168, "东": 169, "临": 170, "乡": 171, "关": 172, "合": 173, "城": 174, "孙": 175, "尣": 176, "幺": 177, "支": 178, "比": 179, "毛": 180, "泽": 181, "無": 182, "生": 183, "臣": 184, "辶": 185, "道": 186, "镇": 187, "黃": 188, "|": 0, "[UNK]": 189, "[PAD]": 190}